How Cropin is democratizing AI - that's Agricultural Intelligence - without field hardware

(Pixabay )

A sustainable food supply chain is essential for the growth of civilization. However, the business of growing food faces growing input costs and low margins. Cropin was a pioneer in digitizing the food supply chain for food processors in India, using this extensive experience and data to democratize crop intelligence for millions of farmers globally. 

I spoke with Cropin Chief Marketing Officer Sujit Janardanan and Head of Cropin AI Labs Praveen Pankajakshan about their important work in bringing sustainability to the food supply chain. Janardanan has been in the B2B technology space for the last 24 years and was previously Head of Marketing for Google Cloud in India. Pankajakshan has been applying AI and machine learning to various problems in energy, healthcare, and agriculture for a couple of decades. He explains:

Our end state vision is can we digitize every single farm on this planet and make it productive and sustainable?  We want every user to have access to their own agricultural intelligence so farmers can ask a simple question like, ‘Hey, based on the weather that you expect to see, what should I do based on where my farm is currently?’  The system should be able to give them an answer so that anyone can get better at improving productivity.

An essential component is a new crop knowledge graph that joins data from different sources to create digital twins for farmers, food processors, and governments. They have digitized over 30 million acres for 7 million farmers in over 130 countries. This helps discover insights buried across the enterprise's ERP, CRM, and supply chain tools and up-to-date information gleaned from satellite imagery using science-based and data-driven models. 

This knowledge graph spans 500 crops and 10,000 varieties built with trillions of data points spanning crop inputs, weather, plant diseases, and yields. The company claims the Cropin Intelligence models deliver consistently high accuracy of 80-90% to decision-makers across the food supply chain. 

Starting with better data

Cropin was founded in 2010 by CEO Krishna Kumar, who was concerned about the rise in farmer suicides fueled by poor harvests. The initial focus was on selling digital tools to farmers directly, but that business model did not scale. Within a year, they found traction with the Cropin Cloud Platform, which helps the food industry digitize its supply chains. Previously, field representatives visited the farms and recorded the expected and actual yields and sales on pen and paper. 

Digital tools helped simplify this process for farmers, field reps, and headquarters planning teams. The food processors bought access to the service. They gave it to farmers for free since it could help food processors identify potential shortages or bumper harvests and adjust their purchasing and processing activities accordingly. Over the years, Cropin expanded to other markets and worked with government planners to help align policies, incentives, and guidance to improve food security and resilience.  

Janardanan said the food industry started asking about integrating data from ERP, CRM, and satellite imagery to improve their planning. This led to the Cropin Data Hub, which helps firms collect, clean, and contextualize data from different sources. 

Once this was in place, Cropin launched an AI practice to enable predictive intelligence. Janardanan says:

That's where you started building models for solving things like identifying the land usage and clearly classifying the land, whether it's cultivation purposes, its forest, or so on. Then, it went to crop classification or identification. Then we started solving for things like yield estimation, crop health, pest and disease infestation, forecast, water stress identification, deforestation mapping, and many others. The primary goal is to help the field staff or the farmer with much more predictive intelligence to act faster than they are probably doing today, where the amount of uncertainty and the rate of change is only increasing alarmingly.

Cutting out the hardware

Some larger farms in developed countries are using drones and sensors to help improve farm intelligence. However, that approach does not scale, particularly for farmers in India and other developing countries who cannot afford to set up an array of sensors and the communication infrastructure to glean insights. 

So, Cropin looked at how satellite imagery could help fill these gaps. For example, recent innovations in multi-spectral imagery analysis tease apart individual crops, disease infestation, water damage, and estimated yields from the unique color combinations seen from space. Janardanan explains:

At the end of the day, it's the farmer who's paying for pretty much everything in the value chain, and hence it's a very cost-conscious industry. Second, agriculture is run on a planet scale, which means that scale is definitely a big challenge for us all. So, if we have to be cost-conscious and need something that can scale, we need to look at the best way to bring in data that can help us solve both of these, and that's where the answer was satellite data. The good part for us is that satellite technology only improves every passing year. It makes it so much more possible, unlike sensors and drones, which are expensive to buy, purchase, deploy, maintain, scale, and standardize. Yes, we can bring in IoT, sensor, and drone data, but that's not the way to go for agriculture.

Combining science and data models

Pankajakshan said another essential insight was the need to combine science-based models with data-driven models to improve accuracy. The science models look at plant varieties and their environmental interaction to predict how the crop will respond to different conditions. He says:

So, there might be traditional heirloom varieties, which might react to the environmental conditions much differently than the current ones or certain other specific varieties which have been developed for, let's say, drought resistance. So, all those things are factored into account. So, when we actually calibrate our models, we require a very minimal amount of data. Because the calibration is done really at the genetic level, the outputs, the yield estimation, and everything is pretty accurate. It just needs that minimum number of samples, which is the advantage of science-based models.

Other data-driven models can provide insight into areas where minimal scientific intervention is required. Since Cropin has been digitizing farmlands for over a decade, they have extensive information about different crops and varieties and how they respond to various climatic conditions. This can help to see and observe how the different sowing and harvesting windows are changing according to the climatic conditions. 

Cropin has also been overlaying this data with satellite data to match the seasonality of each crop with satellite imagery to create the specific training data set for building new AI and machine learning models.  Pankajakshan explains:

It's actually a very dynamic system. Just knowing where crops are growing is quite challenging. So we try to solve that problem first and then identify the crops. And then we overlay it with multiple different situations like stresses, biotic stresses, abiotic stresses, weather-climatic-related stresses, and other things.

Building a digital thread

One essential component of a digital twin is connecting representations of things from different data sources for various types of users. Farmers want to know the best varieties and inputs for predicted weather, food companies want to contract out the appropriate ingredients for their needs, and governments want to ensure food for their nations. 

Cropin has invested significant research and development efforts in correlating the naming conventions and the semantic links associated with entities buried in various data sets. The resulting knowledge graph is a living, breathing translation layer that grows with the discovery of new insights and connections. Pankajakshan notes:

Knowledge is always going to be dynamic, so we keep improving that because earlier knowledge of the crop cycle, the sowing pattern, and the right time of sowing is actually emerging and changing as climate changes. So, as that changes, we need to also ensure that our internal knowledge systems also change.

An essential challenge in building more scalable digital twins in any domain lies in automating the process of building out the ontologies and taxonomies for connecting the different ways of describing things across different data sets, tools, and types of users. Pankajakshan explains:

There are some relationships ontologies and taxonomies that we have already internally defined. But this is really a complex thing; I find that there are some new things coming in every other day that we have not accounted for. And we have to keep updating it. Not all of it can actually be automated. So sometimes, we need to engage with the subject matter experts or field scientists whom we work with as collaborators. We also have subject matter experts internally. 

We also tap into knowledge bases, like Wikipedia. We are creating a centralized data lake, which taps into some of the wiki entries for relationships and graphs. We want to see how we can best standardize those things because the more important question is not only us creating our own ontologies and terminologies and relationships, but also bringing in that external knowledge.

They are also developing some language models to help automate aspects of the process. But Pankajakshan says they are proceeding cautiously since AI tends to hallucinate when interpreting long-tail knowledge. He says:

If you look at the knowledge distribution, it can perform well on what is commonly available. However, where the knowledge that is being asked and inquired is very minimal, which we call longtail knowledge, it starts to hallucinate and make up its own answers. So, we are very careful about how we use the language model itself. We use our internal knowledge systems to enhance the language models. For example, how do you stabilize these systems based on this external information? So, those are things which we are also currently exploring as research ideas and such applications.

A developing field

Cropin is not alone in building tools to distill satellite imagery and supply chain data for agricultural insights. New competitors include companies like CropX, AgroScout, and Agritask. But Cropin does have a head start, a  global reach, and a comprehensive platform. 

Even though these companies compete in many markets, Janardanan says they are excited to collaborate with diverse players to democratize crop intelligence and build a more sustainable, profitable future for all. Opportunities for collaboration include:

• Sharing Data, Strengthening Innovation: New entrants often lack the vast datasets needed for robust AI tools. Pooling resources and sharing anonymized data, can fuel innovation across the board, benefiting everyone.
• Joint Problem-Solving: Combining diverse perspectives fosters creativity. Partnering on specific challenges, like optimizing water usage or combating new pests & diseases, can lead to solutions faster than individual efforts.
• Standardization & Interoperability: Fragmented data formats hinder progress. Collaboration on data standards and platform interoperability can unlock seamless information exchange, benefiting farmers, enterprises, and governments.
• Joint Advocacy & Knowledge Sharing: Collaborating on policy advocacy or joint knowledge-sharing platforms can create a more supportive environment for sustainable agriculture.
• Joint Technology Development: Joint R&D initiatives can tackle complex challenges like disease prediction or automated yield forecasting. Collaboration can accelerate breakthroughs and benefit all.

Janardanan concludes:

The roadmap for the future is how we can democratize access to agricultural intelligence. It shouldn't matter where you are in the value chain, who you are, or what role you play. You should be able to ask a question or query the universe of agricultural knowledge and find an answer. That's where we are headed. The good part is that technologies like generative AI make that journey a lot more possible than it was a year back. 

The second one is whether we can build an ecosystem platform where we see more standardization and less diversification because we have enough diversity to deal with this challenge at the core agriculture layer. We hope we can solve the diversity problem and hence the problem of making everything work together at the technology layer.

My take

Farmers worldwide are struggling with the impacts of climate change, increasing costs, and new environmental regulations that make it harder to turn a profit. It seems like every other week, another group of farmers is making the news for larger-scale protests to bring attention to the challenges they face in one of our most essential industries. I hope Cropin and other companies like it can help ensure farmers can continue to feed us while earning an honest living. 

At the same time, this will not be an easy process. The interests of large food conglomerates that fund these platforms can be at odds with the farmers who grow our food. Cropin and its competitors need to tread cautiously in creating a process to #acceleratetrust between farmers, processors, and farm agencies. It will be essential to include stakeholders from all three as equals as they build out their tools and AI models to ensure all voices are heard.